FAQ: Mercury Playback Engine, CUDA, OpenCL, Metal, and what it all means

Report · Jan 04, 2011

Mercury Playback Engine ExplainedMercury Playback Engine is the given name for a group of improvments that is been built into Premiere Pro. See Adobe documentation on the topic here: Mercury Playback Engine (GPU Accelerated) renderer.

Those improvements includes:

- 64-bit application

- multithreaded application

- processing of some things using CUDA
- processing of some things using OpenCL

- processing of some things using Metal

All versions of Premiere Pro (since CS5) have the first two of these: 64 bit and multi-threaded applicaton.
Since then, GPU Acceleration was added, enhancing Mercury Playback Engine performance.

CUDA support on Nvidia GPUs came first. Then, OpenCL and Metal processing for the Mercury Playback Engine came along.
For current versions of Premiere Pro, you need at least 2 GB VRAM for HD. 4GB for 4K for the Mercury Playback Engine to function as specified.
See System Requirments.

The official and up-to-date list of the cards that provide CUDA, OpenCL, and Metal processing features is here.

Here's a list of things that Premiere Pro can process with CUDA, OpenCL, and Metal:

- some effects

- scaling - (alternate link)

- deinterlacing

- blending modes

- color space conversions

One set of things that Premiere Pro's Mercury Playback Engine doesn't process: encoding and decoding.

That said, two new options for GPU accelerated encoding have been added recently, both of which use different tech than the Mercury Playback Engine:

GPU acceleration is available for decoding and encoding of H.264 and HEVC formats with certain Intel GPUs using Intel Quick Sync. Info here.
GPU acceleration is also available for encoding of H.264 and HEVC formats regardless of CPU type. Info here.
The three technologies do work in concert, with default presets for hardware acceleration set accordingly.

Note that whether a frame can be processed by CUDA, OpenCL, or Metal depends on the size of the frame and the amount of VRAM on the graphics card. This article gives details about that. Error Compiling Movie errors are often at the root this issue.

Processing with CUDA, OpenCL, or Metal doesn't just mean that things are faster. In some cases, it can actually mean that results are better, as with scaling. See this article for details.

If you don't have a supported GPU, you can still use Premiere Pro; you just won't get the advantages of processing with CUDA, OpenCL, or Metal.
- For that, use Mercury Playback Engine Software Only mode. This mode is also useful as a troubleshooting tool to check any anomalies with effects or visuals. The drawback is that it is a lot slower than working with a supported GPU.

An article on the Premiere Pro team blog based on the information and questions in this forum thread has been posted, please check that out.

Notes

The author of this post is no longer working at Adobe, so it needs to be updated by my team. The various Premiere Pro Team Blog links to this article still function as of June 2020, however, their removal is likely imminent. This info needs to be captured before its removal, as it contains critical information for editors. We ask for your patience as this gets fixed.
This article covers information about the Mercury Playback Engine GPU Acceleration. Please do not confuse this information with GPU accelerated exporting technology. More info on GPU accelerated exporting here.
As of June 2020, on macOS, CUDA processing for the Mercury Playback Engine been deprecated for Premiere Pro. Use Metal now.

Report · Nov 04, 2011

KlausKi wrote:
Dave_Gish wrote:
Huh? I usually add 2-3 effects on every clip, and with CUDA there's no need to pre-render. Yellow bars might as well be green. The program windows runs smooth at 1080p, full playback resolution.
Yes, presumingly this is the case when you're using a CUDA card that's supported by Adobe. But as you may have read from my previous posts, I am not using a CUDA card that's supported by Adobe yet.
Dave_Gish wrote:
To be clear, Adobe is saying that if you enable CUDA with a card that's not on their list, then they're not going to support you if you have a problem. However, since many people are doing this, there's a lot of information on the web to help users resolve CUDA problems on their own.
Can you please point me to this piece of information in the official documentation?
I'm a customer, not an information scout. So: I'm paying, Adobe is supposed to deliver. No-one else, and not vice versa.

If you read my previous posts, you'll notice I'm using a GTS450, which is an unsupported card.

You won't find any mention of the unsupported method in official documentation, precisely because its unsupported.

If you're not comfortable hunting for information, then I suggest using one of the supported cards.

Report · Nov 04, 2011

my wish would be that these yellow and red bars in the timeline would turn green so that I don't just see a still in the Preview window as soon as I add two or three effects to a clip.

Just a note, background rendering is a nice to have feature, but in theory, if you're leveraging everything that CUDA gives you, you should be able to avoid rendering most, if not all of the time.

Yellow should still play; red is the only thing that requires rendering here for previewing, and that only happens when you're not using CUDA accelerated effects. Just making sure you know how to pick the realtime ones: look for "Filter effects by type" in the following link for how to quickly filter for CUDA effects in the Effects Panel:

http://help.adobe.com/en_US/premierepro/cs/using/WS774370E7-584E-40e2-B2E0-B33A0228BF02.html

Report · Nov 05, 2011

Thanks, Will, for sheding light on this and sharing your knowledge with me!

OK then, in the meantime I have also read this blog post by Todd...

So I'd like to amend my wish from stating "yellow and red" to "anything that Premiere Pro recognizes as not being able to render in realtime in an acceptable quality, i.e. frame rate".

Full resolution rendering would not necessarily be a feature required for automatically rendering previews in the background, because most of the time the preview window is just a small part of the full screen. If speed is an issue, Premiere Pro could provide an option to render previews in a two-pass process: Low resolution rendering first, then high resolution rendering when there's no more clip/effect left to be rendered in low resolution. Parallel processing should be performed, i.e. Premiere Pro should spawn multiple background rendering threads, one per idle CPU core.

Just like Windows 7 is able to calculate a system performance rating according to some benchmark tests, Premiere Pro should be able to do the same at installation time in order to being able to estimate which of the clips/effects in particular whould require pre-rendering - depending on the built-in hardware.

Report · Nov 05, 2011

Dave_Gish wrote:
If you read my previous posts, you'll notice I'm using a GTS450, which is an unsupported card.
You won't find any mention of the unsupported method in official documentation, precisely because its unsupported.
If you're not comfortable hunting for information, then I suggest using one of the supported cards.

Yes, Dave, you are right about the impact of my suggestion on CUDA cards...

My suggestion tends to be kind of a general one, addressing all users of Premiere Pro, even those using ATI graphics cards, on-board graphics cards or other graphics cards... It would generally be a good thing if Premiere Pro would (optionally) automatically pre-render preview clips in the background, e.g. while the user is editing Premiere Pro effect settings, or creatively thinking about what to do next, or answering the phone... That's plenty of spare time for the machine to perform pre-rendering without requiring the user to actively trigger this process.

And for the other issue: you are right again, Dave, because that's exactly my point:

The documentation is misleading. And this fact needs to be corrected. So other users, who can't afford spending their time hunting for pieces of information they even don't surmise that they would be in need to know about, won't get into the same expensive and time consuming mistaking situation I got.

Todd's blog post reads: "The official and up-to-date list of the cards that provide the CUDA processing features is here:" ... This sentence is - as we all know by now - technically wrong. There are other NVIDIA cards on the market providing "the CUDA processing features" indeed. Adobe can't possibly track all available cards on the market, so a "list of the cards that provide the CUDA processing features" provided by Adobe is basically irrelevant.

BTW: I just went to the shop today to undo the GTX 580 update. I have re-ordered a Sparkle GTS 450 Passive and have them build-in that card again. Now I'm aspiring the same great results you get with your GTS 450

Report · Jan 18, 2012

Hi! I read your blog, but still there is something I do not understand. I hope this is the right thread to ask.

I have Premiere Pro CS 5.5 Installed with a processor i7 950 with 12MB in RAM, some months ago I upgraded my graphics card to an NVIDIA QUADRO 2000 (previously I had and FX1800) as I wanted to have a more fluid editing and rendering process.

But I have not seen at all the difference.
Premiere, does recognize the card, as know I have enabled the GPU option in the software, but that's it.

A few weeks ago I got a Laptop, which has a newer processor i72820QM, but slower one and a good Nvidia card GEFORCE GT 525M, but not one approved for Mercury GPU acceleration and 8MB in RAM.

Amazinlgy, editing with 2 or 3 layers of AVCHD multicamera video works much better, running more fluid than the first one and seems to render faster.

So, I got a software to see what is the ussage of the GPU during different processes (gpuz) and I found out that none of them, neither the laptop or the PC are using the GPU at all during my editing process.

Is this correct? I would expect to see some usage here, but no, GPU usage stays between 0 and 1% all the time.

I kept doing tests and found at that when encoding to MPEG2 for DVD creation. I sent a 27 seconds clip, using Adobe Media Encoder and it took 4.5 minutes to finish and NO GPU usage at all. Then I sent exactly the same clip with same output requirements through the direct export of premiere pro and it took only 15 seconds, and the GPU was used at 96%.

That is a great difference!! but using the Premiere Export, instead of the Encoder Queue is not as useful.

So how can I get Premiere to use the GPU like that as much as it cans? Shouldn't it be using it all the time?

Best regads,

Report · Mar 10, 2012

I'm new to PPCS5.5 and need to get a GPU card that supports CUDA - before I buy a Quadro 4000 (I'm on a Mac) to replace my outdated ATI Radeon HD 3870

- there won't be any change/compromise in my ability to use openGL in Photoshop, correct? I'm assuming that a better video card is better accross all software? I just need to make sure that my performance in PS is as good (or better)...

many thanks!

Report · Mar 10, 2012

What CUDA does...

http://blogs.adobe.com/premiereprotraining/2011/02/cuda-mercury-playback-engine-and-adobe-premiere-p...

Report · Mar 10, 2012

jason dewey wrote:
I'm new to PPCS5.5 and need to get a GPU card that supports CUDA - before I buy a Quadro 4000 (I'm on a Mac) to replace my outdated ATI Radeon HD 3870
- there won't be any change/compromise in my ability to use openGL in Photoshop, correct? I'm assuming that a better video card is better accross all software? I just need to make sure that my performance in PS is as good (or better)...
many thanks!

I'm not a heavy Ps user, but I am a heavy Ae user. The only diff I noticed after replacing my GE Force 8800 with the Quadro is that it hobbled some of my GPU accellerated effects, such as Optical Flares, Shapefifter and Freeform. I've read that the issue is with the drivers for 10.6.8, and that when I upgrade to Lion, I should be able to use Freeform and Shapeshifter again. I can use Optical Flares now, but I have to turn GPU accelleration off, or Ae will crash every time. Freeform and Shapeshifter both crash Ae upon invoking the effect with 10.6.8 and the current Quadro driver. So, take that into your consideration. You might hear similar stories in Ps forums.

Report · Mar 10, 2012

thanks Jim! I'm also on 10.6.8 so I guess I should pose the question on a PS forum as well. I can't imagine that it would impact most PS tasks, but I do work with very large layer files and ocassionaly some render-intensive filters like warp, liquify, and lens correction etc.

Report · Mar 12, 2012

I'm getting the idea that people shouldn't go into the GPU/Cuda business expecting to encode quicker right?

But on the website of NVIDIA they have this video in which they proudly present the fact that encoding in AME is so much faster due to the use of the GPU, see: http://www.nvidia.co.uk/object/adobe_PremiereproCS5_uk.html

What am I missing?

Sebastiën

Report · Mar 12, 2012

It's not the encoding that's faster; it's the rendering.

Exporting has two major phases: 1) rendering the frames and 2) encoding the rendered frames to an output format.

Premiere Pro uses CUDA to accelerate (1) but not (2).

Report · Mar 14, 2012

Clear Todd, thanks.

Just as a side question then: If you render out the project in PP before going to AME and then check the "use rendered files" box in AME before hitting export, does that have any influence (negative or positive) as opposed to not rendering the project in PP and having AME do both the rendering and the encoding?

Thanks a lot,

Sebastiën

Report · Mar 14, 2012

Influence is either None in the rare case where source material and export are identical, or Negative in the frequent cases where FPS, dimensions, codec and bitrate change. Usually it is best not to use Preview files.

Report · Mar 14, 2012

Seb B, the only advantage to what you decribed is a bit of time saving when you Export. If you dont mind rendering the timeline with lots of various effects before the export process, and you choose a high qulity codec for your Preview filesa as well as Export files, you can save some time on the back end when you export. This is becuase at this time all the rendering of effects is done and a part of your preview files. At this point all PrP (using AME) is doing is a simple rebuild of each frame as it creates your exported master file/

Report · Mar 14, 2012

Thanks guys, I'll let AME do everything from now on then ("hello 30 hour encoding times again...", haha), makes sense if you want to achieve the highest possible quality.

Report · Mar 15, 2012

lasvideo wrote:
the only advantage to what you decribed is a bit of time saving when you Export. If you dont mind rendering the timeline with lots of various effects before the export process, and you choose a high qulity codec for your Preview filesa as well as Export files, you can save some time on the back end when you export.

Except that this doesn't always work in Pr Mac 5.5.2. I routinely render my sequence using ProRes422 Previews, and when I export to ProRes422, it takes just as long to export as it did to render, no matter what settings I use. I've tried them all. Some times the exports are fast; other times, I might as well not have rendered my Sequence. I can find no predictable pattern to when it works or not.

This should be a huge timesaver in Pr... more than "a bit of time" saved. Too bad it always isn't. Hope it gets fixed in CS6.

Report · Mar 15, 2012

Jim -"I might as well not have rendered my Sequence. "

Excprt with sequences that have lots of effects or exceed your computers ability to playback, you have to render your sequence to watch it and make editorial decisions. Some clinets like to see it to.

Report · Mar 15, 2012

True. But, most of the time, I work alone, without benefit (ahem) of client in the room. My projects often go through dozens of iterations, and I might be making one or two little tweaks before I FTP an mp4 off to my waiting clients. That's the case where there's no advantage to rendering first, especially if the Export is going to ignore my renders.

What's odd about it is often my computer is fast enough to play all the effects without rendering, but the exporting is 10x the actual run-time of the spot.

Report · Mar 15, 2012

"True. But, most of the time, I work alone, without benefit (ahem) of client in the room. My projects often go through dozens of iterations, and I might be making one or two little tweaks before I FTP an mp4 off to my waiting clients. That's the case where there's no advantage to rendering first, especially if the Export is going to ignore my renders."

Lucky duck

"What's odd about it is often my computer is fast enough to play all the effects without rendering, but the exporting is 10x the actual run-time of the spot."

It sounds like something is definately not right with that situation.

Report · Mar 15, 2012

Well, it took a few decades and advances in technology to get where I could work alone. If I had to buy all the digital tape decks, it would have been much harder to bring my business home. FTP, DG FastChannel, and card/drive media has enabled a paradigm shift. I hardly ever see my clients. Even the new ones, who I got on referral. I have a couple I still haven't met face to face.

Something is not right for sure with my exporting. I wish I could figure it out. I've tried all the standard repair and troubleshooting tactics. Tried the new Mac OS (10.7.3). I haven't reinstalled Pr yet, but I suppose that's on the list of things to try.

Report · Mar 15, 2012

Do you have the Quadro 4000? I know it doesnt do encoding, but if you have any accelerated effects (etc) it helps. And even Colorista 2 is GPU accelerated. It definately helps when rendering in AE. Ive recently seen some benchmarks that make that point dramatically.

Report · Mar 15, 2012

Yes, that's the reason I get such great real-time performance. I try to mainly use accellerated effects. Hard to get around not using Magic Bullet Looks, though. It's not on the list (yet... hopefully some day).

I haven't noticed any speed improvements in Ae, which I use a lot. In fact, the Quadro is not compatible with a few plugs (ShapeShifter and FreeForm Pro), but I've been able to work with OpenGL on, with fewer problems. So, I guess maybe that does give some speed to Ae.

Report · Mar 15, 2012

Have you run the AJA test to determine your raid speed?

http://www.aja.com/support/kona-pc/kona-pc-3g.php

AJA KONA System Test Version 1

A Windows utility for measuring system performance with AJA KONA Video Capture cards. The application includes disk drive speed tests and video data copy (DMA) speeds. The disk speed tests differ from standard disk I/O performance applications in that they specifically test the system under conditions typically encountered with video capture, playback and editing. Note: most of the disk performance tests are generic and don't require an AJA KONA Video Capture card. The DMA performance tests do require a KONA card and AJA KONA Driver software to operate.

Report · Mar 15, 2012

Oh, yeah. I have a 8-SAS drive 6Gbs ATTO RAID. AJA System Test reports between 1500-2100 MB/s reads and writes on the Sweep Video File Sizes test.

And as I mentioned, the export slowness comes and goes, even in the same project. I'm working on a three minute promo, and the exports can be between 7 and 25 minutes. Even 7 seems slow, on a rendered sequence.

I have other weird issues with Pr as well. Some times, I'll scrub the CTI, and there is no updating whatsoever in the Record window. Some times it's 1 FPS. Some times it's thousands of FPS. I'm just not detecting a cause/effect relationship. I've read some other threads here lambasting Pr for having memory leaks. I don't even know what that really means, and consequently if it applies to my situation. If it weren't for the CUDA processing, I'd frankly probably be working in Media Composer or FCP7. The pros outweighs the cons for Pr at this point.

Report · Mar 15, 2012

Yeah, thats great performance for a raid.

So we both know the MPE thrives on fast disk access, lots of memory, CUDA cards and processing power. Any of these seem like a suspect?

I guess another factor is the format/codec of your timeline and your choice of Export format/codec.